Atlas Research Note
NVIDIA Moves Agent Safety Into the Infrastructure
Ben Bajarin · September 28, 2026
Atlas Research Note | September 28, 2026 Our Read How much work enterprises hand to agents comes down to how much authority they can safely delegate. Our security work has focused on identity, permissions, and the ability to stop or reverse an unsafe action. NVIDIA's Open Agent Safety Platform adds a very concrete question to that list: do those controls still work when the computer running the agent can't be trusted anymore?